Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!ames!apple!mips!zaphod.mps.ohio-state.edu!rpi!batcomputer!cornell!uw-beaver!rice!sun-spots-request
From: lm@sun.com (Larry McVoy)
Newsgroups: comp.sys.sun
Subject: Re: SPARC Context switching times
Keywords: SunOS
Message-ID: <5680@brazos.Rice.edu>
Date: 9 Mar 90 08:15:48 GMT
Sender: news@rice.edu
Organization: Sun-Spots
Lines: 110
Approved: Sun-Spots@rice.edu
X-Refs:  Original: v9n72
X-Sun-Spots-Digest: Volume 9, Issue 80, message 6

In article <5539@brazos.Rice.edu> gja@mullian.ee.mu.OZ.AU writes:
>X-Sun-Spots-Digest: Volume 9, Issue 72, message 7
>
>The subject almost says it all. Does anyone know where I can find info. on
>this. The issue is that we are interested in how fast an interrupt routine
>will be activated after the actuall interrupt line is activated. Advice on
>which FM to read would be appreciated.

Well, I'm not sure how you want to count this: each process goes down into
the kernel to send the sig, back out of the kernel, and then back until
the signal is sent back from the other process.  I think you can believe
that you went back and forth 10000 times.  Both processes read a wall
clock time duration so how you want to compute the "context switch time"
is beyond me.  Anyway, this may help you get a rough idea.  On a SS1 I get
something like 710 usec each in the printout which I think means 350usec /
switch.  Furthermore, running the program under rusage shows mean 10011
voluntary switches but another 2400 involuntary switches.  That cuts it
down to 564usec / switch which gives 282 usec each.  It's on the order of
300 usecs is my best guess.

Remember - Sun's have "contexts" so rapidly switching between two
processes is faster than switching between 100.  It doesn't scale
linearly.  Furthermore, measuring ctx switch time this way isn't very
meaningful - the processes are tiny, the pages are always there in memory,
the cache is probably full and hitting like crazy etc, etc.

#include <sys/types.h>
#include <signal.h>

pid_t other;
int caught;

void handler() { caught++; }

main()
{
	struct sigvec sv;
	int parent;

	sv.sv_handler = handler;
	sv.sv_mask = 0;
	sv.sv_flags = 0;
	sigvec(SIGUSR1, &sv, 0);
	sigblock(sigmask(SIGUSR1));

	if (!(other = fork())) {
		other = getppid();
		parent = 0;
	} else {
		parent = 1;
	}
	start();
	while (caught < 10000 && (kill(other, SIGUSR1) != -1))
		sigpause(0);
	kill(other, SIGUSR1);
	stop();
	printf("This time is real time, divide by two for ctx sw time\n");
	ptime(caught);
	if (parent)
		wait(0);
}

/* 
 * statistics code
 */
#include "stdio.h"
#include "sys/types.h"
#include "sys/time.h"

static struct timeval t1, t2;

error(s)
    char* s;
{
    perror(s);
    exit(1);
}

start() { gettimeofday(&t1, 0); }
stop() { gettimeofday(&t2, 0); }
ptime(bytes) { ptransfer(bytes, &t1, &t2); }

ptransfer(bytes, t0, t1)
	long bytes;
	struct timeval *t0, *t1;
{
	struct timeval td;
	double s, bs;

	tvsub(&td, t1, t0);
	s = td.tv_sec + (td.tv_usec / 1000000.);
#define	nz(x)	((x) == 0 ? 1 : (x))
	bs = bytes / nz(s);
	printf("%d in %.4g secs, %.6g / sec, %.6g millisec each\n",
		bytes, s, bs, s / (bytes / 1000.0));
}

tvsub(tdiff, t1, t0)
	struct timeval *tdiff, *t1, *t0;
{

	tdiff->tv_sec = t1->tv_sec - t0->tv_sec;
	tdiff->tv_usec = t1->tv_usec - t0->tv_usec;
	if (tdiff->tv_usec < 0)
		tdiff->tv_sec--, tdiff->tv_usec += 1000000;
}

What I say is my opinion.  I am not paid to speak for Sun, I'm paid to hack.
    Besides, I frequently read news when I'm drjhgunghc, err, um, drunk.
Larry McVoy, Sun Microsystems     (415) 336-7627       ...!sun!lm or lm@sun.com